The opponent is not “not trustworthy enough.” The opponent is a false mission: it looks like service, and does not look like failure when it fails. The patient is still on a phone, the train is still moving, the ledger is still being written—only the purpose has already changed. Without trustworthiness, a charge is empty talk; without a charge, trustworthiness is a means. Mission contains trustworthiness. Trustworthiness is not another highest concept standing beside it.

Without trustworthiness, a charge is empty talk; without a charge, trustworthiness is a means.

To advance into the age of mission-bound software is not another industrial slogan. It is to bind software back to one inspectable charge.

01I. The vital systems have already been handed to a stranger

Software has become the operating system a modern state actually uses. That is not a metaphor. A blackout, a stoppage of trains, of clinics, of clearing, today often begins when a program stops. Unix, C, databases, compilers, the cloud and large models turned mathematics and engineering into executable form. Finance, communications, the grid, railways, cars, factories, hospitals and city government no longer have a spare nervous system that can run outside software. Artificial intelligence is still writing cognition and action into that nerve.

We would not hand a city to a doctor who cannot show the grounds of a diagnosis and cannot be questioned—however busy, expensive or new he looks. Software is that stranger: dependence can no longer be stopped, yet few can read its grounds, and still fewer can force it to halt when it says “continue.”

Therac-25 is not nostalgia. It is a prototype of a transfer of trust—from a physical interlock to a program’s guarantee, then from that guarantee to a place where no one can put a question. What we face now is not one accelerator, but thousands of potential, hidden cases of the same kind.

02II. Behind fifteen trillion, what is missing is not scale. It is the charge

In 2025 China’s software and information-technology services booked RMB 15.48 trillion in software business revenue, up 13.2%; profit RMB 1.88 trillion; more than 11 million people employed. IT services RMB 10.64 trillion; software products RMB 3.24 trillion; industrial software products RMB 333 billion. By any measure this is a large and living industry.

Two other figures on the same sheet matter more. Foundational software: RMB 214.6 billion, 1.39%. Information-security products and services: RMB 223.5 billion, 1.44%. Together, less than three percent of the whole. The structural numbers do not say the industry has failed. They say the next layer has not been built.

Classical software engineering has built the most formidable quality apparatus in the history of human engineering: languages, types, compilation, testing, static and dynamic analysis, review, configuration, security, formal verification, DevOps. This essay holds that in full respect. But its core premise is that correctness is defined relative to a specification, and it pushes “whether the specification itself is right” outside its own border. A system can be made exactly as required and still execute something that was never charged to it. A gate is proved to accept only a red permit—the proof is perfect—when what was wanted was authorization. That is the specification gap.

Fifteen trillion answers scale. It has not answered: of that fifteen trillion, how much software can produce an argument of behavior, showing that it is still executing the one thing it was charged to do.

03III. The opponent is not “not trustworthy enough.” It is a false mission

The old question was written: are there defects? Today’s real question has already gone far beyond defects. Code can be right, the supply chain can show a list, a model can post a score, an agent can execute. A list is not evidence. A declaration is not a mission.

A false mission looks like service. It replaces the charge with sentences that cannot be falsified—“empowerment,” “intelligence,” “serve the people.” Serve the people is not a banner here. It is a named list: patients, passengers, dispatchers, depositors, soldiers. Prosperity outside that list may be a real business. It cannot impersonate this charge. Software without evidence of mission cannot claim to serve the people.

Mission-bound software is not a product category cut by industry, nor *mission-critical* in English. It is software bound, in a particular relation of charge, by five things at once:

Miss one, and the object falls back to a product or a vision. Trustworthiness here is an epistemological organ: an argument others can take away and try to falsify. It is not a synonym for trusted computing. Trusted computing supplies a hardware root of trust; dependability answers whether the system is still working; trustworthiness answers how you know. The three complement one another. None replaces mission. Mission sits on top.

04IV. Generation is accelerating. Inspection is falling behind

The U.S. National Institute of Standards and Technology’s official account is that vulnerability submissions grew 263% from 2020 to 2025. In 2025 it finished enriching nearly 42,000 vulnerabilities, 45% more than in any year in its history. It ran faster than it ever had. It still lost. On 15 April 2026 the National Vulnerability Database it runs abandoned, as a matter of policy, complete analysis of every vulnerability. The question of trust is no longer only whether software should be trusted. It is whether we still have the capacity to judge whether software is still executing its mission.

Generation and verification are asymmetric. Large models push the cost of generation down and the speed of generation up. That is a productivity revolution. It does not automatically supply inspection. The stronger the model, the cheaper the generation, the higher the share of verification in total cost. Artificial intelligence is both a tool and a new object: when a model writes the code, when an agent performs the act, fluency can precede understanding, and action can precede correction.

“Let AI accept AI” does not go through in type. Statistical judgment answers whether something looks right; deductive judgment answers whether a rerun yields the same result; institutional judgment answers who has the authority to decide that it is right. No accuracy, however high, lets statistics climb into the institutional tier. Persons must remain in three seats: legislation, boundary-drawing, judgment. To put the strongest judgment in front of a monitoring screen is this generation’s most expensive waste.

05V. The supply chain has stretched the trusted computing base until it burst

There is almost no truly independent program left. A system depends on hundreds, thousands, tens of thousands of components. The trusted computing base is enlarged without limit by the supply chain. Verizon’s 2026 *Data Breach Investigations Report* shows exploitation of vulnerabilities as an initial access path rising to 31%, overtaking stolen credentials for the first time in nineteen years. The 2025 report in the same series shows incidents involving third parties rising from 15% to 30%. An earlier essay must be corrected: 48% is the share of breaches involving ransomware in the 2026 report, not the share involving third parties.

IBM’s 2025 *Cost of a Data Breach Report* gives a global average of USD 4.44 million per incident, and USD 10.22 million in the United States. 4.44 million is the first fall in five years. It cannot be written as “the cost only grows.” The explanation of the fall is worth more than the absolute: AI on the defensive side works; the U.S. record high is tied to regulatory fines. In the same report, 97% of compromised organizations had not deployed AI access control.

The cost of software error was never only a security incident. Faults, downtime, a wrong financial entry, a wrong medical dose, a supply-chain accident, all enter the same ledger. This is not a quality problem for software firms alone. It is an infrastructure problem for the whole digital economy.

06VI. Understanding the world and taking action can still execute a purpose that was never written down

Palantir is often written as the specimen of “understand the world and take action.” The other face is what is to be contested. Its ontology contains objects, relations, action types and dynamic security; the official formula is a fourfold integration of data, logic, action and security. It does not lack action, nor permission. What it truly lacks is verification and evidence: was this action right? On what grounds? Can it be reconstructed? Are the rules consistent with one another?

Without verification and evidence, understanding the world can still execute a purpose that was never written down. The next generation cannot stop at “what is this, what can be done, who may do it, what was done.” It must still answer: why is it allowed; where did the ground come from; is this decision still on that one charge; what, in the end, happened, and can the evidence be taken away and falsified by someone else.

The materials can be written: data, knowledge, ontology, rules, software, AI, verification, execution, evidence. The rope is another thing. However complete the materials, without charge, forbidden zone, binding, trustworthiness and account, the system can still execute whichever purpose comes more readily to hand.

07VII. Make a system, or bind a purpose

Mission-bound software is the object. Mission-bound software engineering is the engineering that makes that object hold. The two must not be mixed.

The core is only one thing:

Without an explicit charge, inspection faces a more convenient objective function; however strict, it is still a false mission. Without still-executing, the definition stays on paper. Without inspection that can be falsified, the charge is a declaration and trustworthiness is empty talk.

The essential distinction is not a few more documents, nor a change of cadence. Classical software engineering makes a system; mission-bound software engineering binds a purpose. It is not a course opened beside software engineering. It is software engineering’s next orientation, after the charge itself became the problem. It inherits the engineering capacity of software engineering. It does not inherit “delivery equals success.” If the criterion is not changed, the base only delivers a false mission at greater expense. If the base is refused, mission is only a vision that changed its words.

Four layers must be set up apart. Connotation: the engineering of a charge; false mission as opponent; inspectability as success; persons in legislation, boundary-drawing and judgment. It is not correctness 2.0, not security under a new skin, not trusted computing, and not a Knowledge Operating System. The apparatus can be changed. The connotation cannot. Method: principles before steps, steps before tools; the forbidden zone before capability; define, align, achieve; generation and inspection fail independently; intensity tied to consequence. Generation covers breadth; deduction covers depth. Theory: the object is a relation of charge; the problem is still executing this one thing; the criterion is how you know—plus the specification gap, the generation–verification asymmetry, the transfer of trust, conditional guarantee. Composability remains unsolved. Engineering system: the inherited base, plus principles, objects, layers, links and replay. A Knowledge Operating System is the name of the apparatus, used to make the core computable. It is not a shipped operating system. Prototypes in a repository can test some mechanisms. They cannot be taken as evidence that a specification has already been realized.

A system can be claimed. The discipline has not yet formed.

08VIII. Not a niche of the security industry. An entry condition

Cybersecurity cannot solve this alone. Security asks whether a property can be kept under malice; most problems of trustworthiness have nothing to do with malice: wrong knowledge, conflicting rules, undeliberated defaults, silent drift of behavior. Ken Thompson said you cannot trust code you did not create yourself. When code is increasingly generated by machines, the sentence goes deeper: not only was it not created by our hands; even whether it was understood cannot be determined.

Large models will not solve it by themselves. Classical software engineering cannot solve it alone. What is needed is a new synthesis: make knowledge and rules first-class computable citizens; make actions and traces recordable, inspectable, revocable; let evidence arise in ordinary running; let trustworthiness walk from chip to action, and not stop at an isolated node.

This is not another isolated “security-software industry.” It is a layer cut across every software domain: from trusted chips and foundational software and a Knowledge Operating System, to ontologies, mission-bound AI, mission-bound agents, industry software, certification and services. Grading and evidence formats are the premise of pricing, certification and trade. The time window for mission-bound agent infrastructure is the tightest. To bring build provenance into the technical criteria of Xinchuang is, for now, the most operable institutional move.

Taking 2025’s RMB 15.48 trillion of software business revenue as the base, 1% penetration is about RMB 154.8 billion; 5% about RMB 774.2 billion. This is a scenario calculation, not a market forecast. Even as a thin value layer inside the whole software industry, it is enough to form a space of hundreds of billions—even trillions—of yuan. What matters more than cutting out a slice of revenue is unlocking AI value that has already arrived and has not yet been cashed. The value a mission creates may far exceed the revenue it captures.

09IX. A strong country is not scale. It is the power to define a mission

We cannot be content to remain users of software. A large software country means scale. A strong software country contains successive layers: foundational theory, core systems software, standards, ecology, infrastructure, and the next software paradigm. The first layers are catch-up. The last is a leap. Catch-up inside an existing paradigm is hard, because the barrier is ecological lock-in; a leap at a point of paradigm change is possible, because the new paradigm has no existing ecology that must be migrated. Mission-bound software is a real point of change—it changes the definition of the problem. To recognize the point of change, and to occupy a place in it, are two different things.

What is most valuable is to define the mission mechanism by which AI enters the real world. When AI moves from generation to action, what decides how far it may be used is no longer capability alone, but whether the action can be trusted. That mechanism is immature everywhere. China’s industry knowledge can be formalized only by China. The power to define is also a duty: a standard that cannot be independently inspected, whoever defines it, contradicts itself. The right stance is not to seize a standard, but to make something better and let it be inspectable.

China already has a software industry on the order of fifteen trillion yuan, a vast manufacturing system, one of the world’s largest bodies of industrial application, fast-growing AI, a complete software chain, and 11 million software workers. That scale is both an advantage and a constraint: any scheme that only experts can use cannot spread at this scale; the strongest part of that corps must be turned from a false prosperity toward industrial and defense-industrial software. What is missing is not the base. What is missing is the strategic capacity to organize that base into next-generation mission-bound software infrastructure, and people willing to work for many years on problems still unsolved.

10X. What this generation must still answer, and what is not yet undone

Look back: Unix answered how an operating system should be organized; C, how to stay close to the machine and remain portable; Lisp, how a program treats symbols and itself; the Web, how information crosses a boundary; open source, who may change and who may inspect; cloud and mobile, how software becomes a remote capability; Git, how distributed collaboration stays consistent; LLVM, how compiler infrastructure is reused. Large models are answering how software should be produced. What has not been answered is where, after this leap in production, verification and evidence are to be placed.

This generation must still answer: how software can truly understand the real world; how to define, align and achieve a charge; how to set connotation, method, theory and engineering apart, without pretending the discipline already stands; how knowledge becomes part of computation; how AI uses knowledge under a mission; how an agent acts under a mission; how complex software can have verifiable behavior; how every critical decision can be explained, traced and audited. None of these yet has a mature answer. That is the difficulty, and the reason it is worth the work.

What must most be avoided is letting anyone think the problem is already solved. Modeling cost is the chief practical risk of the whole route; it has not been shown to be solved. Composability still depends on engineering discipline, not on an architectural guarantee. Expressiveness and decidability are theoretically constrained. There may be a class of systems whose performance requirements are simply incompatible with a complete mission argument. Standardization of knowledge representation, the actual holding of AI’s boundary, the conflict between keeping evidence and keeping privacy, all remain. Formal compliance replacing a real mission, absorption into a platform so that no independent industry forms, the closing of a window in international standards, can all still happen. These are not disclaimers. They are the real terrain of the route. A system that knows its own border is more trustworthy than one that does not. A route is the same.